By early 2026 a wave of research was measuring and forecasting AI's effects on the labor market. This report, published by Anthropic's Economic Research team on 5 March 2026 as a web page and a 17-page PDF, does not claim to settle the question. Its stated aim is to set up a measurement framework before large effects appear, so that later analyses are less post hoc, and to test it against the data available so far. It opens with a caution drawn from earlier work: a well-known estimate that about a quarter of US jobs were offshorable was followed by healthy growth in most of those jobs, and studies of industrial robots and of the China trade shock still disagree about their employment effects (Introduction). The report finds "limited evidence that AI has affected employment to date" (Introduction).
This note follows the PDF; the web page has the same text with different footnote numbering and an appendix PDF that is linked but not in this library's sources.
What the report does
Exposure measured at the task level
The report follows a common design in labor economics: compare outcomes for workers who are more and less exposed to a technology, so that economy-wide shocks such as the business cycle cancel out (Counterfactuals). Exposure is defined at the level of tasks, not whole jobs; the report's example is that AI can grade homework but not manage a classroom, so teachers count as less exposed than workers whose whole job can be done remotely.
Three data sources are combined (Measuring exposure):
- O*NET, the US Department of Labor's database listing the tasks of about 800 occupations.
- Anthropic Economic Index usage data: classified samples of Claude conversations. The measure uses the two most recent datasets, covering August and November 2025 (PDF footnote 5).
- Eloundou et al. (2023), which scored each task with a value $\beta$: 1 if an LLM alone could make the task at least twice as fast, 0.5 if it would need additional tools or software built on the LLM, and 0 otherwise. Their scoring assumed an LLM with a 2,000-word input limit and no access to recent facts (PDF footnote 4).
Theoretical capability and actual use are closely related but not identical. Tasks rated $\beta = 1$ account for 68% of observed Claude usage and tasks rated $\beta = 0$ for 3% (Figure 1 caption); in total, 97% of the tasks observed in the previous four Economic Index reports fall in categories rated feasible ($\beta$ of 0.5 or 1). The gap runs the other way too: Eloundou et al. rate "Authorize drug refills and provide prescription information to pharmacies" as fully exposed, but the authors have not observed Claude performing it. As general reasons why usage can lag theoretical capability, the report names model limitations, legal constraints, specific software requirements and human verification steps.
The observed exposure measure
Observed exposure asks, of the tasks an LLM could theoretically speed up, which are actually being done with Claude in automated, professional settings. An occupation scores higher when its tasks are theoretically feasible, appear with significant usage in the Economic Index, appear in work-related contexts, are carried out in automated rather than augmentative ways (including through the API), and make up a larger share of the job (Measuring exposure). In practice a feasible task counts as covered once it has enough work-related usage; fully automated use gets full weight and augmentative use half weight; and task coverage is averaged to the occupation, weighted by the share of time spent on each task. The mathematical details are in the appendix.
The authors list the judgment calls this involves, such as how to code $\beta$, what counts as significant use, and how much more automation should count than augmentation, and report that the rank correlation of job exposure across many choices is "exceedingly high" (PDF footnote 6).
Outcome: unemployment
The report chooses unemployment as its main outcome because it most directly measures harm to workers: a fall in job postings for one role might be offset by openings in a related one, but an unemployed worker wants a job and has not found one (Prioritizing outcomes). The data are from the Current Population Survey (CPS), in which unemployed respondents report their previous occupation. Treatment is defined by rank: workers in the top quartile of time-weighted task coverage are compared with workers whose occupations have zero coverage, and the cutoff is varied in robustness checks (Initial results). The comparison uses a difference-in-differences framework, which measures how the gap between the two groups changed after ChatGPT's release in late 2022 relative to before.
What it finds
Coverage is far below theoretical capability
By broad occupational category, the $\beta$ measure shows scope for LLM use in most tasks of Computer and Math occupations (94%) and Office and Administrative occupations (90%), but Claude currently covers 33% of tasks in the Computer and Math category (Figure 2 and text). The most exposed single occupations are Computer Programmers at 75% coverage, followed by Customer Service Representatives, whose main tasks increasingly appear in first-party API traffic, and Data Entry Keyers at 67% (Figure 3 and text). At the other end, 30% of workers have zero coverage because their tasks were too rare in the data to meet the threshold; examples include cooks, motorcycle mechanics, lifeguards and bartenders.
Exposure and projected job growth
The US Bureau of Labor Statistics projects employment by occupation for 2024 to 2034. In an occupation-level regression weighted by employment, each 10 percentage point increase in observed coverage goes with a 0.6 percentage point lower projected growth (Figure 4 and text). The authors call this "some validation" that their measure tracks independent analysts' estimates, while noting that the relationship is slight, and report no such correlation for the Eloundou et al. measure alone.
Who the exposed workers are
Using CPS data from August to October 2022, the three months before ChatGPT's release, the top exposure quartile differs sharply from the zero-exposure group (Figure 5 and text). Exposed workers are 16 percentage points more likely to be female, 11 points more likely to be white, almost twice as likely to be Asian, and earn 47% more on average. Graduate degrees are held by 17.4% of the most exposed group and 4.5% of the unexposed group. The key findings add that exposed workers are more likely to be older.
No detectable change in unemployment
Unemployment rates for the two groups since 2016 move together apart from COVID-19, when the less exposed group, more often in in-person jobs, saw a much larger rise (Figure 6). The average change in the gap since ChatGPT's release is "small and insignificant": the exposed group's rate rose slightly, but the effect cannot be distinguished from zero. The appendix checks, as summarized in PDF footnote 9, vary the treatment cutoff from the median to the 95th percentile (the effect is flat or negative in every case), look at workers aged 22 to 25, and replace survey unemployment with Department of Labor unemployment insurance claims; none finds clear effects on exposed jobs.
The report also states what the design could detect (Initial results). From the confidence interval of the pooled estimate, a differential rise of about 1 percentage point would be detectable, a figure the authors call a ballpark that will change with new data. Two benchmarks illustrate the scale: if every worker in the top 10% of coverage lost their job, unemployment in the top quartile would rise from 3% to 43% and aggregate unemployment from 4% to 13%; a doubling like the 2007 to 2009 recession (5% to 10% nationally) would take the top quartile from 3% to 6%, which should also be visible. Because the estimate is differential, a rise in unemployment that hit all workers equally would not be attributed to AI.
Slower hiring of young workers
Brynjolfsson et al. (2025) had reported a 6 to 16% fall in employment in exposed occupations for workers aged 22 to 25, mainly through slower hiring; the range reflects different counterfactuals (PDF footnote 10). This report finds the unemployment rate of young workers in exposed occupations flat, but notes that slower hiring need not show up as unemployment, since many young people are labor-market entrants with no listed occupation and may leave the labor force instead (Initial results). It therefore uses the CPS panel to measure the monthly rate at which 22 to 25 year olds start a job they did not have the previous month, split by the exposure of the occupation entered (Figure 7).
The two series diverge visibly in 2024. Entry into the least exposed occupations stays at about 2% per month, while entry into the most exposed falls by about half a percentage point. Averaged over the post-ChatGPT period, the job-start rate into exposed occupations is 14% lower than in 2022, an estimate the authors describe as "just barely statistically significant." There is no such drop for workers over 25. The authors list alternative explanations: young people not hired into these jobs may be staying in existing jobs, taking different jobs or returning to school, and surveys may mismeasure job transitions. The web page records a correction on 8 March 2026: Figure 7 had swapped the labels of the two groups.
Limitations and open questions
The authors state these limits in the text and discussion:
- One platform's usage. Coverage is measured from Claude traffic only; the measure can incorporate other usage data and other countries (PDF footnote 3), but this version does not.
- A dated capability measure. The Eloundou et al. scores reflect LLM capabilities as of early 2023 and could be updated (Discussion).
- Unemployment is one channel. The framework "won't capture every channel" through which AI could affect work (Introduction). Changes in wages, hours or job content within an occupation are not measured.
- Rank-based treatment. The analysis assumes the most exposed workers are affected first; if capabilities advance quickly, an absolute threshold might be more informative (Initial results). Other work suggests that effects depend on whether exposure covers all of a job's tasks or only some (Gans and Goldfarb, 2025; Hampole et al., 2025, as cited).
- Weak hiring evidence. The young-worker result is barely significant and open to the alternative explanations above.
- Planned updates. The authors intend to revisit the analysis as usage data accumulate and suggest studying recent graduates with credentials in exposed fields next (Discussion).
Research questions
- Does the job-start gap for young workers persist and grow? The 14% figure is barely significant. Re-estimating it with later CPS months, and with administrative hiring data rather than survey responses, would show whether it is a stable effect or noise.
- Does coverage measured on other AI platforms rank occupations the same way? Observed exposure uses Claude data only. Computing the same measure from another provider's usage data, or from survey data on AI use at work, and comparing occupation rankings would test how platform-specific it is.
- How much does the half weight on augmentation matter for the results? The report says rankings are robust to such choices but the unemployment estimates are not shown for each variant. Re-running the difference-in-differences under alternative weights would show whether conclusions depend on them.
- Do wages or hours move before unemployment does? The report chooses unemployment as the outcome of most concern. Applying the same exposure groups to CPS earnings and hours would show whether effects appear first in channels the main analysis does not track.
- Does the projected-growth relationship hold up in realized employment? The 0.6-point slope compares exposure with BLS projections. Comparing exposure with realized occupational employment in later years would test whether the measure predicts actual change.
- Are customer service occupations a leading case? The report notes that customer service tasks increasingly appear in automated API traffic. Tracking employment and job starts for that occupation separately would show whether high automated coverage precedes measurable change.
How it connects
- Anthropic Economic Index report: Learning curves, published three weeks later, cites this report when it notes that coding tasks are moving from Claude.ai to the API and suggests that this migration may signal "more imminent transformation of work" for the associated jobs.
- The usage data come from the Anthropic Economic Index series; the report cites the January 2026 report by Appel et al. (Anthropic Economic Index report: Economic primitives) and Handa et al. (2025), which introduced the index (The Anthropic Economic Index). It also cites Tamkin and McCrory (2025) (Estimating AI productivity gains from Claude conversations).
References
- Massenkoff, M., and McCrory, P. (2026, March 5). Labor market impacts of AI: A new measure and early evidence. Anthropic. https://www.anthropic.com/research/labor-market-impacts. PDF: https://cdn.sanity.io/files/4zrzovbb/website/2b5bbaf2c1eb81dbf6e6fb813c1a24e35a64d376.pdf. Footnote numbers in this note follow the PDF.
- Task- and occupation-level coverage data: https://huggingface.co/datasets/Anthropic/EconomicIndex (PDF, Data availability).
- Studies cited above only as the report describes them (Brynjolfsson, Chandar and Chen, 2025; Eloundou et al., 2023; Gans and Goldfarb, 2025; Hampole et al., 2025) are listed in the report's references and were not read for this note.